Trading with Asymmetric Volatility Spillovers*
Angel Pardo** and Hipolit Torro**
Universitat de València
Preliminary Draft, June 2003
* CICYT project BEC2000-1388-C04-04 and the Instituto Valenciano de Investigaciones Económicas (IVIE) providedfinancial support. We are grateful for the comments and suggestions of G. Llorente and V. Meneu. The usual caveatapplies.** Professors at the Department of Financial Economics at the University of Valencia. Corresponding author HipolitTorro, Faculty of Economics, University of Valencia, Avda. dels Tarongers s/n, 46022 Valencia (Spain), Tel.: 34-6-38283 92; Fax: 34-6-382 83 70, E-mail: [email protected]
1
Trading with Asymmetric Volatility Spillovers
Abstract
This article studies the dynamic relationships between large and small firms by using Volatility
Impulse-Response Function of Lin (1997) and its extensions, which takes into account the
asymmetric structures on volatility. The study reveals that bad news about large firms can cause
volatility in both large-firm returns and small-firm returns. Furthermore, contrary to the previous
evidence, bad news about small firms can also cause volatility in both kinds of firms. After
measuring spillover effects, different trading rules have been designed. There is evidence that these
rules provide very profitable strategies, especially after bad news coming from its own and the
“other” market. These results are of special interest for practitioners because of its implications for
portfolio management.
Keywords: volatility spillovers, asymmetric volatility and large and small stock exchange indices
2
Trading with Asymmetric Volatility Spillovers
1. Introduction
Unexpected shocks between large and small companies have attracted the attention of both
academics and practitioners because of its implications for portfolio management and asset pricing.
A great number of papers have shown that large-firm returns can be used to forecast small-firm
returns, but not vice versa. In this way, Conrad, Gultekin and Kaul (1991) find that volatility
surprises are important to the future dynamics of their own returns as well as the returns of smaller
firms. However, the converse is not true. This unidirectional causality agrees with the “directional
asymmetry” found in McQueen et al. (1996). They show that the lead-lag relationship between
large and small portfolio returns exists only after unexpected positive shocks in large stock portfolio
returns because unexpected negative shocks are updated immediately.
There are several hypotheses that have tried to explain the existing cross-correlation between large
and small stock returns. The first one is focused on the non-synchronous trading effect. Boudoukh
et al. (1994) and Chelley-Steeley and Steely (1995) find that non-synchronous trading can be an
important determinant of cross-correlations. In stark controversy, Lo and Mackinlay (1990) and
Conrad, Kaul and Nimalendran (1991) indicate that this effect could account only for a small part of
the observed portfolio returns serial autocorrelation. The second hypothesis relates the significant
cross-correlation to the differential quality of information caused both by large and small firms and
to differences in response of them to general economic and firm-specific factors (Yu and Wu,
2001). Finally, the last hypothesis blames the more quantity of information produced from large
companies as the most important reason for the existence of the unidirectional lead-lag relationship
from large to small stock returns (see Chopra et al., 1992 and Badrinath et al., 1995).
Although relationships between large and small portfolio returns are very well documented in the
literature, volatility spillovers between them have not been studied enough. When variances and
covariances are applied to study the dynamic relation between large and small firm returns, it is
important to differentiate between asymmetric volatility and covariance asymmetry. The first one
refers to the empirical evidence that stocks returns are more volatile in bearish than bullish markets
while the second one helps to explain the former (see Bekaert and Wu (2000)).
3
Volatility asymmetry first appeared in the financial literature with Black (1976) and Christie (1982).
The explanation they put forward is based on the “leverage effect hypothesis”: a negative return
increases financial leverage, causing the volatility of the equity’s rate of return to rise. However, it
seems that the leverage effect is too small to fully account for this phenomenon (Christie (1982) and
Schwert (1989)). Another explanation is often referred to as “volatility feedback effect”: if the
market risk premium is an increasing function of market volatility, an anticipated increase in
volatility raises the required return on equity, leading to an immediate stock price decline
(Campbell and Hentschel (1992)).
Which of both competing explanations is the main cause of asymmetric volatility has been an open
question over years. Kroner and Ng (1998) shed more light on this topic by documenting significant
asymmetric effects in both the variances and covariances. In particular, bad news about large firms
can cause volatility in both small-firm returns and large-firm returns. Moreover, the conditional
covariance between large-firm and small-firm returns tends to be higher following bad news about
large firms than good news. Following this line, Bekaert and Wu (2000) provide a general empirical
framework to examine volatility by differentiating between the two competing explanations and by
examining asymmetric volatility at the firm and at the market level. They find evidence that the
volatility feedback effect is particularly strong when the conditional covariance between market and
stock returns responds more to negative than to positive markets shocks.
In volatility symmetric structures, it is not necessary to distinguish between positive and negative
shocks, but with asymmetric structures the Volatility Impulse-Response Function proposed by Lin
(1997) change with the shock sign. Therefore, this methodology can be especially useful for
obtaining information on the second moment interaction between related markets.
The main objective of this paper is to go deeply into volatility spillovers between large and small
firms by studying the impulse-response function for conditional volatility. It is important to point
out that, as far as we know, this is the first time that large-small firm portfolio relationship is
studied by using Lin’s methodology and its extensions. The study of volatility spillovers taking into
account the Volatility Impulse-Response Function can be very helpful in designing trading rules
based on the inverse relationship existing between expected volatilities and expected returns.
Furthermore, we use stock market indices on large and small liquid stocks instead of portfolios.
This fact has two clear advantages for practitioners: they can take signals directly from market
indices quotations, so it is not necessary to build portfolios and, the implementation of the trading
rules can be notably reduced due to the existence of derivative contracts on the large stocks index.
4
The rest of the paper is structured as follows. Section 2 introduces econometric framework and
formulates our empirical model. Section 3 presents the data and discusses the main empirical
results. In section 4, trading rules based on volatility spillovers are designed and computed. The
final section summarises the main results.
2. The Econometric Framework
2.1. The Means Model
As this paper mostly addresses modelling volatility rather than returns forecasting, a two-step
estimation procedure is followed. First, a model in means is estimated and then the residuals of this
model are taken in the second step as an input to estimate the conditional variance. To clean up any
autocorrelation behaviour, the following vector error correction model (VECM) is estimated:
∑∑
∑∑
=−
=−−
=−
=−−
+∆+∆++=∆
+∆+∆++=∆
p
jtjtj
p
jjtjtt
p
jtjtj
p
jjtjtt
SbLazcS
SbLazcL
1,2,2
1,2122
1,1,1
1,1111
εα
εα
(1)
where Lt and St refers to the logarithm of the large and small stock indices respectively, zt-1 is the
lagged error correction term of the cointegration relationship between Lt and St;
ijijiii ba,d,,c and α for i =1,2 and j=1,…,p, are the parameters to estimate, p is the lag of the VECM.
The VECM model is estimated by Ordinary Least Squares applied equation by equation (see Engle
and Granger (1987) and Enders (1995)). The residual series of this model, ε1t and ε2t, are saved and
they will be used as observable data to estimate the multivariate GARCH model. This two steps
procedure (see Engle and Ng, 1993 and Kroner and Ng, 1998) reduces the number of parameters to
estimate in the second step, decreases the estimation error and allows a faster convergence in the
estimation procedure.
5
2.2. The Covariance Model
The number of published papers modelling conditional covariance is quite low compared to the
enormous bibliography on time-varying volatility. One consequence of this lack of studies in
covariance modelling is that asymmetry receipts in volatility are directly extended to the
multivariate setting. Because of the cross effects generated in each multivariate GARCH model, the
natural extension in asymmetry modelling from a univariate to a multivariate setting can have
unexpected effects among all the elements of the covariance matrix. The consequences of this
extension are unclear because there is no evidence enough on how asymmetries behave in the
covariance. The most common case of volatility asymmetry in stock markets is the negative one,
where unexpected falls in prices increase more the volatility than an unexpected increase in prices
of the same amount. Engle and Ng (1993) analyse different asymmetric volatility models; they
show that the asymmetry depends not only on the sign but also on the innovation size. That is, the
asymmetry, if it exists, is clearer when unexpected shocks in prices are important. These authors
propose a battery of tests to verify the importance and sense of the asymmetries. They obtain
evidence for the Glosten et al. (1993) model where a dummy variable is included in a GARCH(1,1)
taking value 1 when the previous innovation is negative.
Multivariate asymmetric GARCH allows for spillover in volatility between large and small firm
portfolios. Furthermore, the cross relationships existing in multivariate modelling allows, for
example, for the small firm portfolio to be sensitive to the large firm portfolio volatility asymmetry
although no asymmetries exist in the small firm portfolio volatility. These kind of cross
relationships can have several consequences in the large-small covariance dynamics, especially in
periods of high volatility.
Kroner and Ng (1998) study asymmetries following the Glosten et al. (1993) approach in a
multivariate setting. This is the most common method for introducing asymmetries in multivariate
GARCH modelling in finance. Kroner and Ng (1998) adopt a structured approach, similar to
Hentschel (1995) nesting the most common covariance models1. Under this framework, model
selection is made easier by testing restrictions and it will allow choosing the right multivariate time-
varying covariance avoiding ad hoc selections. After applying the above mentioned specification
1 The four most widely used models are: (1) the VECH model proposed by Bollerslev et al. (1988), (2) the constant
correlation model, CCORR, proposed by Bollerslev (1990), (3) the BEKK model of Engle and Kroner (1995) and (4)
the factor model proposed by Engle et al. (1990).
6
test, it was picked up the asymmetric extended BEKK model. This model has the following two-
dimensional compacted form:
G'GA'ABH'BC'CH '1t1t
'1t1t1tt −−−−− +++= ηηεε (2)
where C, A, B and G are 2 × 2 matrices of parameters, Ht is the 2 × 2 conditional covariance, εt and
ηt are 2 × 1 vectors containing the shocks and the threshold terms series, see below. So, the
unfolded covariance model is written as follows:
′
+
′
+
′
+
′
=
−
−−−
−
−−−
−
−−
2221
12112
122
12112
111
2221
1211
2221
12112
12
12112
11
2221
1211
2221
1211
122
112111
2221
1211
22
1211
22
1211
22
121100
gggg
.gggg
aaaa
.aaaa
bbbb
h.hh
bbbb
ccc
ccc
h.hh
t
ttt
t
ttt
t
tt
t
tt
ηεηη
εεεε
(3)
Where cij, bij, aij, and gij for all i,j = 1,2 are parameters, ε1t and ε2t are the unexpected shock series
obtained from equation (1). η1t = max [0,−ε1t] and η2t = max [0,−ε2t] are the Glosten et al. (1993)
dummy series collecting a negative asymmetry from the shocks and hijt for all i,j = 1,2 are the
conditional second moment series.
2.3. Asymmetries Analysis
Covariance asymmetry analysis is carried out in two steps. First, a misspecification test on
asymmetries filtering is conducted before and after estimating the asymmetric covariance model.
Second a graphical analysis of news impact surfaces2 and the Asymmetric Volatility Impulse-
Response Functions (AVIRF) is displayed.
The robust conditional moment test of Wooldridge (1990) is applied to test how the Glosten et al.
(1993) modification to the multivariate GARCH models cleans the asymmetries in the conditional
2 A “news impact surface” is defined as the relationship between each conditional second moment (or a function of
them) and the last period pair of shocks holding past conditional variances and covariances constant at their
unconditional sample mean.
7
covariance matrix. This test enables the identification of possible sources of misspecification in the
model, and is robust to distributional assumptions (see also Brenner et. al (1996)). The generalized
residual is defined as ijtjtitijt h−= εευ for all i,j = 1,2, which is the distance between the
covariance, or variance, news impact surface and its T -consistent estimator. Using the same
misspecification indicators as Kroner and Ng (1998), the Wooldridge (1990) robust conditional
moment test is computed. Kroner and Ng (1998) suggest the use of three kinds of indicator
variables to detect misspecification of the conditional covariance matrix. These indicators try to
detect misspecification caused by shock signs ( )( 011 <−tI ε and ( ))012 <−tI ε , the four quadrants
sign combinations ( )( 00 1211 >> −− tt ;I εε , ( )00 1211 >< −− tt ;I εε , ( )00 1211 <> −− tt ;I εε ,
( ))00 1211 << −− tt ;I εε and the misspecification induced because of the cross effect of shock sings
and shock sizes ( )( 0112
11 <−− tt I εε , ( )0122
11 <−− tt I εε , ( )0112
12 <−− tt I εε , ( ))0122
12 <−− tt I εε .
The Volatility Impulse-Response Function (VIRF) is a useful methodology for obtaining
information on the second moment interaction between related markets. The impulse-response
function for conditional volatility is defined in Lin (1997) as the impact of an unexpected shock on
the predicted volatility, that is
[ ])( dg
vech 3 ,
tt
tst,s
|HERεε
ψ∂
∂= + (4)
where Rs,3 is a 3 × 2 matrix, s=1,2,… is the lead indicator for the conditioning expectation operator,
Ht is the 2 × 2 conditional covariance matrix, )',()( dg 22
21 t,t,
,tt εεεε = , ψt−1 is the set of conditioning
information. The operator ‘vech’ denotes the operator that transforms a symmetric N × N matrix
into a vector by stacking each column of the matrix underneath the other and eliminating all
supradiagonal elements.
In symmetric GARCH structures it is not necessary to distinguish between positive and negative
shocks to obtain the VIRF, but with asymmetric GARCH structures the VIRF must change with the
8
shock sign. The VIRF for the asymmetric BEKK model3 is taken from Meneu and Torro (2003) by
applying (4) to (2),
>++=
= +−
+1s) ½(1s
n,1sn,s Rgba
aR (5)
>++=+
= −−
−1s) ½(1s
n,1sn,s Rgba
gaR (6)
where +n,sR ( −
n,sR ) represents the VIRF for positive (negative) initial shocks and where c is a 3 × 1
parameter vector and a, b and g are 3 × 3 parameter matrices4. The AVIRF asymptotic distribution
is obtained straight away from VIRF results appearing in Lin (1997).
3. Data and Empirical Results
3.1. Data
The data used in this study has been provided by the Sociedad de Bolsas, which manages the most
important indexes of the Spanish Stock Exchanges. Specifically, the data used in this study consists
of daily closing values of the IBEX-35 index and the IBEX-Complementario from January 2nd of
1990 to June 28th of 2002.
The IBEX-35 index is composed of the most liquid 35 securities quoted on the Spanish Joint Stock
Exchange System of the Four Spanish Stock Exchanges (Madrid, Barcelona, Bilbao and Valencia).
The IBEX- Complementario index is composed of the securities included in the Sectorial Indexes of
3 The probability distribution of shock signs is needed to obtain the conditioning information flow at any time. It is
assumed that prob (εt < 0) = ½ and prob (εt ≥ 0) = ½ for all t. Furthermore, shock sign independence over time and
independence between shock signs and squared unexpected shocks is also assumed.
4 Where NN D'B'BDb )( ⊗= + , NN D'A'ADa )( ⊗= + , NN D'G'GDg )( ⊗= + ,
=
100010010001
ND is a duplication matrix,
=+
10000½½00001
ND is its Moore-Penrose inverse and ⊗ denotes the Kronecker product between matrices.
9
the Sociedad de Bolsas that do not belong to the IBEX-35. 5 It is important to point out that only the
most liquid stock are traded in the Spanish Joint Stock Exchange System. Therefore, this is a
guarantee of liquidity. However, in order to overcome possible problems associated with thin
trading, weekly frequency is used, taking Wednesday closing values or the previous trading day if
the Wednesday is not a trading day. Returns series are obtained by taking first differences in the log
prices.6
3.2. Preliminary Analysis
Figure 1 displays the weekly evolution of the stock indices IBEX-35 and IBEX-Complementario in
the studied period and preliminary data analysis is presented in Tables I, II and III. Table I displays
returns, volatilities and correlation coefficients, year by year through the sample period for both
stock indices, the IBEX-35 (It) and the IBEX-Complementario (Ct). Three facts can be highlighted
from this table. First, there are 4 years (1994, 1999, 2000 and 2002) in which both indices offer a
different sign return but means equality hypothesis can not be rejected. A second appealing fact in
Table I is that after 1992 the IBEX-35 volatility is fairly larger than the IBEX-Complementario
volatility. Variance equality test rejects the null in years 1999 to 2002. Finally, last column shows
that the correlation between both series has decreased as time pass.
Differences in means and variances can be understood as both classes of stocks offering different
sensitivities to risk factors. For example, large companies are more internationalised depending on
global risk factors and small companies risk factors are localised basically in its own economy. This
fact can be seen as a globalisation effect on the Spanish stock market through several global crises
(European Monetary System suffered several crises in the early nineties, Asian crisis in October
1997,…), international strategic positions taken by the most important Spanish companies
(especially in Latin America) and, simultaneously, these companies have begun to be traded in the
most important stock markets in the world.
5 During the studied period, the IBEX-35 and IBEX-Complementario represented the 82% and the 6%, respectively, of
the overall capitalisation of the Spanish Stock Exchange.
6 The common tests of unit roots and cointegration (Dickey and Fuller (1981), Phillips and Perron (1988), and Johansen
(1988)) offered no doubt about this point.
10
From Table II, it can be stated that the pair of financial time series used in this paper offers very
similar statistics. Both have significant skewness, kurtosis, autocorrelation, heteroskedasticity and a
single unit root. Moreover, although equality in means can not be rejected the variances equality
test is rejected. This preliminary result points out that more research is necessary in the covariance
dynamics between both financial time series.
3.3. Estimating the Model
The model in equations (1) and (3) is estimated in a two-step procedure. To take account of the pre-
holiday effect on the Spanish Stock Exchange7, a dummy variable has been also included in the
mean equation. The model for the means is:
∑∑
∑∑
=−
=−−
=−
=−−
+∆+∆+++=∆
+∆+∆+++=∆
p
jt,jtj,
p
jjtj,ttt
p
jt,jtj,
p
jjtj,ttt
CbIaHOLdzcC
CbIaHOLdzcI
122
122122
111
111111
εα
εα
(7)
where HOLt is a dummy variable that equals to one when the next weekly return contains a pre-
holiday.
First, the VECM model in equation (7) is estimated by Ordinary Least Squares applied equation by
equation (see Engle and Granger (1987)). The VECM lag was chosen by maximising AIC criterion.
Table III shows that series are cointegrated being 3 the optimum lag length. Panel (A) in Table IV
displays the estimated coefficients and the residual analysis.
Examination of the speed of adjustment coefficients (α1 and α2) provides insight into the
adjustment process of stock indices towards the long-run equilibrium. For the stock indices to adjust
to the long-run relationship it is necessary that α1 > 0 and α2 < 0 (assuming IBEX-35 to be weak
7 See Meneu and Pardo (2003).
11
exogenous and IBEX-Complementario endogenous) 8. The estimated coefficients have the expected
sign in the IBEX-Complementario (α2 ) equation but it has the opposite sign in the IBEX equation
(α1 ). It can be conclude that the IBEX-35 leads the IBEX-Complementario in the long run.
From Table IV, it can be seen that the pre-holiday dummy coefficients are significant in both
equations. This variable has already been studied in Meneu and Pardo (2003) with daily series, but
it is the first time that it is found significant in weekly series. So it can be inferred that it is a very
important anomaly and it should not be omitted.
The residual analysis in Table IV shows that with the estimated model autocorrelation disappears
but heteroskedasticity remains. Furthermore, in Panel (B) Granger causality tests reject the null
hypothesis. Therefore, there is no causality in any sense in the short run.
Table V and VI display the estimated conditional covariance model and its standardised residual
analysis, respectively. Table V estimates have been computed assuming a conditional normal
distribution for the innovation vector (ε1t,ε2t)’. The standard errors and their associated critical
significance levels are calculated using the quasi-maximum likelihood method of Bollerslev and
Wooldridge (1992) which are robust to the non-normality assumption. Panel (A) displays the
estimates. The low critical significance level obtained for 13 out of 15 parameter estimates reveals
that this model fit very well with the data.9 Restrictions on cross-variance effects and asymmetric
covariance are clearly rejected (see Panel (B)). As a consequence, cross relationships across all
conditional moments and their shocks (symmetric and non-symmetric) cannot be skipped.
Moreover, asymmetries but it self are also significant. Panel (C) shows the estimated persistence to
any shock in the estimated conditional covariance model. The conditional variance for the IBEX-35
has a very high persistence to its own shock. But volatility persistence is a quite common feature in
financial time series.
Table VI displays standardised residual analysis. From this table it can be concluded that
autocorrelation and heteroskedasticity problems have been successfully amended. Finally, Figure 2
displays the conditional second moments evolution overall the estimation period. Both volatility
8 From the cointegration relationship, the error correction term can be written as tttt ICz 21 ββ −−= . For the
existence of a long-run equilibrium relationship, it is necessary that 011 >−tzα and 012 <−tzα . See Johansen (1995),
chapter 8 for more details.
9 The maximum log-likelihood function value obtained in the estimation process was 4460.
12
series have similar patterns but Ibex-35 volatility is almost ever above of the Ibex-Complementario
volatility, especially after 1997 when the globalisation effect on the biggest companies in the IBEX-
35 is particularly important.
3.4. Filtering Covariance Asymmetries
Table VII contains the results of the robust conditional moment of Wooldridge (1990) to test how
the Asymmetric BEKK-GARCH model cleans up the asymmetries in the conditional covariance
matrix. Panel (A) displays the test result when unconditional covariance estimate is used. It can be
seen than asymmetries are very important, especially in the covariance between both series and beta
coefficients. Panel (B) offers the test results once asymmetries are included in the covariance
specification. After this, only one asymmetric pattern seems to remain in the conditional covariance
specification. This is an important result because it means that the GARCH specification is
gathering almost all the possible asymmetries in the conditional covariance matrix: direct and
crossed asymmetries and asymmetries of sign and size in the unexpected shocks. This result is also
a guarantee that the analysis of the asymmetric volatility impulse response function carried out later
is reliable. That is, it will be able to answer the following questions: Are spillovers of volatility
important in the large-small firm system? Are the unexpected negative shocks of large stocks
conditional variance important in the small stocks conditional covariance? And the reverse? Which
market leads the volatility system?
The effect of asymmetric behaviour in conditional beta coefficients is also investigated. Last
column in Table VII contains its robust conditional moment test10. The presence of any asymmetric
effect is clearly rejected in Panel (B). This result shows that conditional beta estimates are
insensitive to volatility asymmetries. This appealing result comes from the fact that a ratio between
two second moments tends to compensate the asymmetric effect if a stable proportion is maintained
between both conditional second moments. This lack of sensitivity is important for portfolio
management based on beta estimates, as it seems unnecessary to consider asymmetries11, therefore
simpler models can be used. It is also important because beta coefficients are a market risk
10 Following Wooldridge (1990), a consistent estimator of the minimum variance hedge ratio is built using the
continuous function property on consistent estimators (see Hamilton (1994), p. 182).
11 Results on symmetric BEKK model show that conditional beta estimates are insensitive to any asymmetry in the
conditional covariance matrix. Results are omitted to conserve space.
13
sensitivity measure and it is shown that conditional beta estimates are insensitive to sign and size
shocks. Furthermore, this result is comparable to the Braun et al. (1995) and Bekaert and Wu
(2000) empirical findings on beta coefficients: while asymmetries are very strong in the conditional
second moments they appear to be entirely absent in conditional betas.
Figures 3-a and 3-b display the unconditional and conditional correlation and beta coefficients,
respectively. It can be appreciated that conditional correlation is quite stable around is unconditional
estimate with a smooth decreasing trend. The conditional beta coefficient has a similar but more
acute pattern. Recovering Figure 2, where IBEX-35 conditional volatility has increased in level
after 1998, it can be inferred that both markets have a weaker relationship than before 1998. So,
diversification strategies are getting an important role in portfolio management.
Figure 4 collects the news impact surfaces for the conditional second moments and the conditional
beta obtained from the asymmetric bivariate GARCH specification.12 It can be appreciated that the
IBEX-35 variance surface shows a clear size asymmetry when opposite signs are registered in both
markets. The IBEX-Complementario variance surface shows a clear sensitivity to its own negative
shocks when positive shocks on the IBEX-35 comes together. Moreover, covariance surface is quite
plane, increasing smoothly as negative shocks in the IBEX-Complementario takes larger values.
Finally, it can be seen that when large and small stock shocks are perfectly correlated, the beta
coefficient is quite stable on its unconditional value (0.67) and it is insensitive to shock size. But
when large cross-signed shocks are allowed (quite possible in this financial system) conditional
betas fall to small values, a very wise result.
3.5. Measuring Volatility Spillovers
Cheung and Ng (1996) propose a no causality in variance test based on the residual cross-
correlation function and robust to distributional assumptions. Causality in variance is of interest to
both academics and practitioners because of its economic and statistical significance. First, changes
in variance are said to reflect the arrival of information and the extent to which the market evaluates
and assimilates new information. As Ross (1989) shows, conditional variance changes are related to
12 Following Engle and Ng (1993) and Kroner and Ng (1998), each surface is represented in the region εit = [-5,5] for i =
1,2.
14
the rate of flow information. So, one way to study how information flow is transmitted between
large and small companies is studying its volatility relationships. Second, the causation pattern in
variance provides an insight concerning the characteristics and dynamics of economic and financial
prices, and such information can be used to construct better econometric models describing the
temporal dynamics of the time series.
Cheung and Ng (1996) no causality in variance test can be viewed as a natural extension of the
well-known Granger causality in mean13. This test is based on the asymptotic distribution of the
cross-correlation function trying to detect causal relations and identify patterns of causation in the
second moment. Panel (A) in Table VIII displays the cross-correlation test for the standardised
residuals obtained from Equation (7). The model is cleaned up of level correlation but there remains
cross-correlation in both senses across squared standardised residuals. After estimating the
covariance model in Equation (3), significant cross-lagged-correlation between standardised
residuals and squared standardised residuals disappears. It is important to stress that cross-lagged-
correlation in both senses exists in squared standardised residuals (Panel A) but they disappear after
introducing the GARCH model (Panel B). So, the AVIRF analysis must be able to exhibit volatility
spillovers across markets.
Figures 5 and 6 present the asymmetric volatility impulse-response functions (AVIRF) computed
following Lin (1997) and Meneu and Torro (2003). Now, it is possible to split the volatility
spillover effect depending on the unexpected shock sign. When unexpected shocks are positive —
Figure 5—, graphical analysis shows that there exist a relatively low volatility spillover from the
small to the large stock index (about 5% of the shock —Figure 5-A), but not the reverse—Figure 5-
F,. If unexpected shocks are negative —Figure 6—, graphical analysis shows that there exist bi-
directional volatility spillovers between both markets. It can be observed that negative shocks in the
IBEX-Complementario have an important effect on its own volatility that takes about 10 weeks to
be absorbed —Figure 6-C, and it is spilled to the IBEX-35 volatility but with a small impact level
(about 7% of the shock —Figure 6-A). On the other hand, about a 5% of a negative shock in the
IBEX-35 volatility is spilled to the IBEX-Complementario volatility —Figure 6-F. By comparing
Figures 5 and 6, we can observe that good and bad news coming from the IBEX-35 have a similar
13 Whether the causality in mean has any potential effect on the test for causality in variance –or vice versa– depends on
the model specification. In a GARCH-M the causality in variance is likely to have a potential large impact on the
causality in mean. As this test can be also used to test no causality in mean both test can be used simultaneously to
improve model specification highlighting the causal relationships both in mean and variance.
15
impact on its own volatility, taking a very long time to die out due to its persistence —Figures 5-D
and 6-D. Finally, the only kind of news items affecting to the IBEX-Complementario volatility are
the negative ones, specially its own negative shocks, taking about 10 weeks to be absorbed—Figure
6-C.
These figures are according to the coefficient estimates of matrix G. First, g11 (−0.0794) and g22
(0.5752) are the coefficients collecting the impact of a negative shock on the IBEX-35 and the
IBEX-Complementario on their own conditional variance, respectively. It is quite clear than only in
the case of the small firm index, the negative volatility asymmetry is relatively important on its
dynamics. Second, g12 (−0.1978) and g21 (0.1895) coefficients reveal than negative asymmetries
spillovers between both markets have a very similar size.14
The empirical results presented here will add evidence against the hypothesis of unidirectional
variance causality from large to small stock portfolios. So the common conclusion of volatility
spillover from large to small stock portfolios (see Kroner and Ng, 1998) may be due to model
misspecification. The AVIRF analysis uncover that any volatility shock coming from small stock
market is important to large stock market but the reverse is only true for bad pieces of news coming
from large stock markets. That is, good news in the large stock markets are not signals for small
stock traders but bad items of news are. Therefore, it can be said that main source of information
comes with bad news coming from any market and it spreads into the ‘other’ stock market.
4. Trading strategies
In order to explore possible consequences to portfolio managers of the uncovered volatility
spillovers across sized portfolios some trading strategies are designed based on them. There are two
competing hypotheses about covariance asymmetries in stock market known as ‘leverage effect’
and ‘feedback effect’. The second one has the empirical evidence and it is based on the existence of
a negative relationship between expected returns and expected volatility.
We have designed a trading rule in order to exploit the inverse relationship between expected
returns and expected volatility once conditional volatility is forecasted. The trading rule consists of
14 Coefficient g12 (g21) measures the impact of a negative shock in the IBEX-35 (IBEX-Complementario) in the “other”
market.
16
selling assets when conditional variance is forecasted to increase and buying assets in the opposite
case. Furthermore, it is possible to exploit volatility spillovers across markets taking signals from a
related market. That is, if an increasing volatility spillover is forecasted then you must sell stocks in
the spilled market. If a decreasing volatility spillover is forecasted then you must buy stocks in the
spilled market. The above section has highlighted a different volatility response depending on the
unexpected shock sign so it is necessary to know the sign of the last item of news in order to
improve volatility forecasting ability.
Table IX displays trading rules designing and its results based on ex post and ex ante volatility
changes are presented in Table X and Table XI, respectively. The period taken to test the
profitability of these strategies is January 2nd, 2001 to June 30th, 2002 with 78 weekly
observations. During this period of time, the IBEX-35 had a clear decreasing trend and the IBEX-
Complementario was quite stable on its initial level. Table IX presents all the strategies designed.
Panels (A) and (B) show how to take positions in the stock market when last item of news is
negative and positive, respectively. In each panel, overall strategies are classified into ‘Direct’ and
‘Crossed’, depending on the market from which signal arises. Finally, if conditional variance is
forecasted to increase a short position must be taken and a long position must be taken in the
opposite case. Strategies A1 to A8 are those taking short positions in the stock indices when a
volatility increase is forecasted in its own or in the spilling market. Strategies B1 to B8 are those
taking long positions in the stock indices when a decrease in its own or in the spilling market
volatility is forecasted. In Table X ex post results are obtained with the estimated model appearing
in Table V and using estimated volatilities for the studied period. In Table XI ex ante results are
obtained by estimating the model each time new weekly returns are known, and forecasting
volatility for the next week, taking stock positions depending on the volatility forecast.
Panels (C) in Tables X and XI display the buy and hold strategies return for stock indices and for
the risk-free investment15. In this period, returns were −30.63% for the IBEX-35, 0.54% for the
IBEX-Complementario and 5.95% for the accumulated risk-free investment. The number of weeks
with positive and negative return is also computed. During this period of time there were 35 (47)
weeks with positive return and 43 (31) weeks with negative return in the IBEX-35 (IBEX-
Complementario).
15 The accumulated weekly Spanish Treasury bill repo rate is taken as the risk-free investment.
17
Before considering profitable a trading rule, transaction cost must be considered. Approximately,
institutional investors trading on the IBEX-35 will require no more than a 0.5% expenses in
transactions costs (commissions, spreads, …). The IBEX-Complementario will require no more
than 1%. Finally, if the futures contract on the IBEX-35 is used instead of the spot index, no more
than 0.1% expenses will be required. Results on the futures contract are not displayed but are
virtually identical to the spot index using the first to delivery contract. So the viable strategies are
those than have a positive return after considering transaction costs of 0.1% in the IBEX-35 based
strategies and 1% in the IBEX-Complementario case.
Tables X and XI show similar results. There only exists one strategy that is profitable ex post but it
is no profitable in the ex ante case, the strategy A8 . All the remaining strategies identified as
profitable or non-profitable agree in both tables16. Hence, the following comments refer to both
tables but after excluding this strategy. Profitable strategies are marked with an asterisk in the case
of positions taken in the IBEX-35 and with two asterisks in the case of the IBEX-Complementario
positions. The profitable strategies are the following: A1 , A2 , A3 , A4 , A5 , B1 , B7 and B8 .
From net returns it is easy to see that the four strategies involving short positions after bad news
( A1 , A2 , A3 and A4 ) have the better performance. So one can conclude that when an increase in
volatility is forecasted after bad news, then a short position must be taken in the stock market. It is
important to stress that strategies based on signals coming from the neighbour market ( A2 and A4 )
are very profitable. It is important to point out that this is a decreasing period for the IBEX-35 and
selling rules will tend to have positive returns. However, this is not true for the IBEX-
Complementario (see Figure 1). It should be remembered that there are 43 out of 78 weeks with
negative return in the IBEX-35, but there are only 31 in the IBEX-Complementario. Therefore, this
is not a hazardous result.
The four strategies with better performance ( A1 , A2 , A3 and A4 ) involve taking short positions in
the stock market. For large financial institutions is possible to take short positions in both the small
and large stock index. In concrete, taking short positions in the IBEX-35 is very easy to any
investor by joining its futures market. So one can conclude that profitable strategies have existed in
the studied period of time.
16 If a risk-free investment position were taken each non-trading week, it should be added at least a 3% of extra return.
18
5. Summary and conclusions
In this article we study the dynamic relationships between large and small firms taking into account
asymmetric volatility and covariance asymmetry. When these types of structures appear, it is
necessary to distinguish between positive and negative shocks. The Volatility Impulse-Response
Functions proposed by Lin (1997) and extended by Meneu and Torró (2003) become especially
useful in this case, since they give information on the second moment interaction between related
markets, and they allow practitioners to design trading rules based on the inverse relationship
existing between expected volatilities and expected returns.
The main result is that there exist volatility spillovers across sized portfolios in both senses after bad
pieces of news. Therefore, bad news about large firms can cause volatility in both large-firm returns
and small-firm returns but bad news about small firms can also cause volatility in both kind of
firms. Ross (1989) demonstrated that variance changes are related to the rate of information flow.
Our results indicate that only bad piece of news contains information, no matter the size of the firm.
After measuring spillover effects, different trading rules have been designed. Specifically, a trading
rule taking advantage of this empirical supported feature is selling assets when conditional variance
is forecasted to increase and buying assets in the opposite case. Furthermore, it is possible to exploit
volatility spillovers across markets taking signals from a related market. Results show that very
profitable strategies exist, especially after bad news coming from its own and the ‘other’ market.
Whether this result is against rational market hypothesis or it can be explained by time-varying risk-
premiums overcomes the objectives of this paper and it is left for further research.
19
6. Tables
Table IReturns, volatilities and correlations
Annualised Returns (%) Annualised Volatilities (%) Correlation (3)
Year IBEX-35 (1) IBEX-Compl. (1) KW Test (5) IBEX-35 (2) IBEX-Compl.(2) Levene Test (6)
1990 -30.40 -41.69 0.00 24.74 28.97 0.04 0.91
1991 13.96 12.55 0.29 15.90 20.40 2.25 0.91
1992 -10.12 -14.89 0.00 20.73 17.41 0.93 0.82
1993 40.08 30.19 0.41 17.19 15.82 0.48 0.83
1994 -13.30 4.38 0.12 21.90 20.64 1.22 0.89
1995 14.00 2.51 0.32 15.45 15.21 0.01 0.88
1996 36.95 30.10 0.39 14.18 10.80 2.89 0.79
1997 34.18 25.95 0.76 20.41 15.57 3.07 0.88
1998 30.44 32.58 0.00 27.70 21.85 2.01 0.76
1999 16.33 -20.68 3.63 24.89 12.54 8.95* 0.74
2000 -24.35 12.18 0.00 25.69 15.58 10.38* 0.64
2001 -7.89 -8.58 0.00 24.92 18.96 5.92* 0.77
2002(4) -46.98 19.31 2.82 21.68 12.15 9.24* 0.68
(1) This column displays the annualised return of the heading index computed from its weekly mean returnin that year as the mean return multiplied by 52.
(2) This column displays the annualised volatility of the heading index computed from its weekly returns inthat year as a sample standard deviation multiplied by (52)0.5.
(3) This column displays the annual correlation between both indices computed from their weekly returns inthat year.
(4) This row displays the results for the period January 2nd to June 28th in 2002.(5) This column displays the means equality test between weekly means known as Kruskal-Wallis.(6) This column displays the variances equality test between weekly variances known as Levene.
Significant coefficients at 95% of confidence level are highlighted with one (*) asterisk.
20
Table IISummary statistics for the data
tI∆ tC∆Mean 0.0012 0.0011
Kruskal-Wallis Test 0.4914 [0.4833]Variance 0.0009 0.0006
Levene Test 20.6246 [0.0000]Skewness -0.4953 [0.0000] -0.4480 [0.0000]Kurtosis 1.1941 [0.0000] 3.1056 [0.0000]
Normality 65.2923 [0.0000] 283.41 [0.0000]Q(20) 33.4216 [0.0303] 56.9280 [0.0000]Q2(20) 85.0268 [0.0000] 85.0268 [0.0000]A(20) 54.55 [0.0000] 54.55 [0.0000]
ADF(4) -0.9852 ⟨-2.5693⟩ -0.7918 ⟨-2.5693⟩PP(6) -0.8443 ⟨-2.5693⟩ -0.5974 ⟨-2.5693⟩
Notes: Kruskal-Wallis statisitic tests the means equality and its p-value appears in brackets. Levene statistictests the variances equality and its p-value appears in brackets. Skewness means the skewness coefficient andhas the asymptotic distribution N(0,6/T), where T is the sample size. The null hypothesis tested is theskewness coefficient is equal to zero. Kurtosis means the excess kurtosis coefficient and it has an asymptoticdistribution of N(0,24/T). The hypothesis tested is kurtosis coefficient is equal to zero. Normality means theBera-Jarque statistic test for the normal distribution hypothesis. The Bera-Jarque statistic is calculatedT[Skewness2/6+(Kurtosis-3)2/24]. The Bera-Jarque statistic has an asymptotic 2χ (2) distribution under thenormal distribution hypothesis. Q(20) and Q2(20) are Ljung Box tests for twentieth order serial correlation in
t,Cε , t,Iε and 2t,Cε , 2
t,Iε respectively and A(20) is Engle (1982) test for twentieth order ARCH; all these tests aredistributed as 2χ (20). The ADF (number of lags) and PP (truncation lag) refers to the Augmented Dickey andFuller (1981) and Phillips and Perron (1988) unit root tests. Critical values at 10% of significance level ofMackinnon (1991) for the ADF and PP test (corresponding to the process with intercept but without trend) aredisplayed as ⟨.⟩ and marginal significance levels are displayed as [.] in the remaining tests.
Table IIIJohansen (1988) tests for cointegration
Lags Null λtrace(r) λmax(r) Cointegration Vectorβ’=(1, β1, β2)
3 r = 0r = 1
17.792.53
15.262.53
1, −0.775, −1.671
95% c. v. r = 0r = 1
15.413.76
14.073.76
Notes: Lags is the lag length of the VECM model in equation (7); the lag length is determined using the AIC.λtrace(r) tests the null hypothesis that there are at most ‘r’ cointegrating relationships against the alternative thatthe number of cointegration vectors is greater than ‘r’. λmax(r) tests the null hypothesis that there are ‘r’cointegrating relationships against the alternative that the number of cointegration vectors is ‘r+1’. Críticalvalues are from Osterwald-Lenum (1992, Table 1). β’=(1, β1, β2) are the coefficient estimates of thecointegrating vector where the coefficient of Ct is normalised to be unity, β1 is the coefficient of It and β2 isthe intercept term.
21
Table IVOLS Estimates of the Error Correction Model and Granger Causality Tests
∑∑
∑∑
=−
=−−
=−
=−−
+∆+∆+++=∆
+∆+∆+++=∆
3
122
3
122122
3
111
3
111111
jt,jtj,
jjtj,ttt
jt,jtj,
jjtj,ttt
CbIaHOLdzcC
CbIaHOLdzcI
εα
εα
PANEL A: OLS Model EstimatesExplanatory variable Depenent Variable
tI∆ tC∆
1−tz -0.0434 (-3.16) -0.0421 (-3.65)
tHOL 0.0088 (2.84) 0.0053 (2.08)
1−∆ tI -0.0597 (-0.91) 0.0582 (1.07)
2−∆ tI 0.0936 (1.42) 0.0163 (0.30)
3−∆ tI 0.0209 (0.32) -0.0744 (-1.38)
1−∆ tC 0.0159 (0.20) 0.0300 (0.46)
2−∆ tC 0.0127 (0.16) 0.1141 (1.78)
3−∆ tC -0.0255 (-0.34) 0.1812 (2.88)
Residual Analysis2R 0.0453 0.076
Log-likelihood 3183.809AIC 3183.868
Q(20) 16.69 [0.6728] 24.36 [0.2267]Q2(20) 74.83 [0.0000] 129.27 [0.0000]A(20) 50.66 [0.0002] 76.99 [0.0000]
PANEL B: Granger Causality Tests03121110 === ,,, bbb:H 0.0548 [0.98]
03222120 === ,,, aaa:H 0.9907 [0.40]
Notes: t-Student values are displayed as (.) and marginal significance levels are displayed as [.]. Q(20) andQ2(20) are Ljung Box tests for twentieth order serial correlation in t,1ε , t,2ε and 2
1 t,ε , 22 t,ε respectively and
A(20) is Engle (1982) test for twentieth order ARCH; all these tests are distributed as 2χ (20). 2R represents theadjusted determination coefficient. Log-likelihood is the maximun value of the log-likelihood function of thesystem. The Akaike Information Criterium of the system appears as AIC. The Granger Causality Test statistichas a F(3,639) distribuition under the null hypothesis.
22
Table V Multivariate GARCH model estimates and restrictions tests
Panel (A). Multivariate GARCH model estimatesG'GA'ABH'BC'CH '
tt'tttt 11111 −−−−− +++= ηηεε
( ) ( )
( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
−−
=
−=
−=
×−=
−
0000005752018950
0000001978007940
0000007788001380
0000000168092870
2500000198021100
0000001348035270
9901083
0000000120000760
8
....
....
G
....
....
B
....
....
A
..
....
C
Panel (B). Testing restrictions on the model
sting cross-variance effects significance (H0: a12=a21=b12=b21=d12=d21). 1.61 106 (0.00)
sting asymmetric variance significance (H0: d11=d12=d21=d22). 1.09 103 (0.00)
Panel (C). Estimated persistence shocks
Half life 11h 22h 12h
t,1ε211
211
211 ½gba ++
69.20
212
212
212 ½gba ++
0.21111211121112 ¼ ggbbaa ++
0.43
t,2ε 221
221
221 ½gba ++
0.25
222
222
222 ½gba ++
2.68222122212221 ¼ ggbbaa ++
0.16
Panel (A) of this table displays the quasi maximum likelihood estimates of the BEKK assuming aconditional normal distribution for the innovation vector (ε1τ.ε2τ)’. Critical significance levels appear inbrackets.Panel (B) displays Wald test restrictions on the covariance model. Critical significance levels appear inbrackets.Panel (C) displays the Half-life as a measure of persistence of any squared shock in each element of the
conditional covariance matrix, approximated with the following formula: )gbaln(
).ln(lifeHalf 222 ½50++
=− ,
using the right coefficients appearing above.
23
Table VISummary statistics for the Standardised residuals
from the VECM-Asymmetric BEKK Model
Dependent Variable
t,t h/ 111ε t,t h/ 222ε
Mean -0074 -0.011Variance 1.029 0.998Skewness -0.5252 [0.0000] -0.4652 [0.0000]Kurtosis 0.9492 [0.0000] 1.7382 [0.0000]
Normality 54.1220 [0.0000] 104.9578 [0.0000]Q(20) 17.77620 [0.6021] 21.7987 [0.3515]Q2(20) 22.1983 [0.3298] 17.2743 [0.6351]A(20) 17.6605 [0.6097] 17.0650 [0.6487]
Notes: Skewness means the skewness coefficient and has the asymptotic distribution N(0;6/T), where T isthe sample size. The null hypothesis tested is the skewness coefficient is equal to zero. Kurtosis means theexcess kurtosis coefficient and it has an asymptotic distribution of N(0,24/T). The hypothesis tested iskurtosis coefficient is equal to zero. Normality means the Bera-Jarque statistic test for the normaldistribution hypothesis. The Bera-Jarque statistic is calculated T[Skewness2/6+(Kurtosis-3)2/24]. TheBera-Jarque statistic has an asymptotic 2χ (2) distribution under the normal distribution hypothesis. Q(20)and Q2(20) are Ljung Box tests for twentieth order serial correlation in t,1ε , t,2ε and
21 t,ε , 2
2 t,ε respectively and A(20) is Engle (1982) test for twentieth order ARCH; all these tests aredistributed as 2χ (20). Marginal significance levels are displayed as [.] overall the tests.
24
Table VIIRobust conditional moment tests
Panel (A). VECM Model.
Generalised residual tests122112 httt −= εευ 11
2111 htt −= ευ 22
2222 htt −= ευ 1112
2121 hhttttbeta −= εεευ
( )00 1211 << −− tt ;I εε 41.1666*** 0.1726 0.8747 305.9997***( )00 1211 >< −− tt ;I εε 50.0959*** 0.0095 0.2497 311.9997***( )00 1211 <> −− tt ;I εε 30.6636*** 0.0633 0.5989 249.9997***( )00 1211 >> −− tt ;I εε 13.3555*** 0.3289 0.5083 55.9999***( )011 <−tI ε 28.5992*** 1.0480 1.0416 61.9999***( )012 <−tI ε 72.3229*** 0.0069 1.7949 278.9998***
( )0112
11 <−− tt I εε 0.4273 2.1365 2.6833 67.4989***( )012
211 <−− tt I εε 0.2633 2.3276 2.8508 66.5590***
( )0112
12 <−− tt I εε 0.3166 2.5870 2.0743 55.6470***( )012
212 <−− tt I εε 0.3065 2.5053 2.0079 55.0708***
Panel (B). VECM- Asymmetric BEKK model.
Generalised residual teststttt h122112 −= εευ ttt h11
2111 −= ευ ttt h22
2222 −= ευ ttttttbeta hh 1112
2121 −= εεευ
( )00 1211 << −− tt ;I εε 0.00914 0.39252 0.2297 0.0368( )00 1211 >< −− tt ;I εε 1.50912 0.04206 1.92339 0.1173( )00 1211 <> −− tt ;I εε 0.00769 0.10958 0.45041 0.1131( )00 1211 >> −− tt ;I εε 0.15527 0.3273 0.17754 0.0885( )011 <−tI ε 7.16526*** 1.05423 3.25917* 0.0004( )012 <−tI ε 1.38256 0.00536 1.48498 0.0459
( )0112
11 <−− tt I εε 0.33422 0.62012 0.25339 1.0712( )012
211 <−− tt I εε 0.09787 0.33743 0.06865 0.6277
( )0112
12 <−− tt I εε 1.00249 0.34513 3.39857* 0.6866( )012
212 <−− tt I εε 0.29722 0.62012 0.25339 0.2370
Panel (A) gives the robust conditional moment test statistic applied on unconditional moment estimates whereh11, h22, h12 and beta coefficient are unconditional estimates of IBEX-35 variance, IBEX-Complementariovariance, its covariance and beta, respectively. Panel (B) gives the robust conditional moment test on theconditional moment estimates, where h11t, h22t, h12t and beta coefficient are the conditional estimates of IBEX-35 variance, IBEX-Complementario variance, its covariance and beta, respectively, obtained from theasymmetric GARCH model. The misspecification indicators are listed in the first column and the remainingcolumns in each panel give the test statistic computed for the generalised residual calculated as the first row ineach panel shows. ε1t-1 is the return shock to the IBEX-35 and ε2t-1 is the return shock to the IBEX-Complementario. The indicator function I() takes the value one if the expression inside the parentheses issatisfied and zero otherwise. All the statistics are distributed as a χ2(1). Test values highlighted with one (*),two (**) and three (***) asterisks are significant at 90%, 95% and 99% of confidence level, respectively.
25
Table VIIICross-correlation in the levels and squares of standardised residuals
Panel (A)Standardised residuals
from theVECM
Panel (B)Standardised residuals
from the VECM-GARCH
Lag k ( )kt,t, ˆ,ˆ −21 εερ ( )22
21 kt,t, ˆ,ˆ −εερ ( )kt,t, U,U −21ρ ( )2
221 kt,t, U,U −ρ
-5 -0.0169 0.0855* -0.0337 0.0267-4 -0.0221 0.0731* 0.0062 -0.0181-3 0.0037 0.1038** -0.0017 0.0153-2 -0.0002 0.0199 0.0041 -0.0231-1 0.0065 0.0921** 0.0329 -0.01350 0.7985** 0.5786** 0.8107** 0.6333**1 -0.0040 0.1173** 0.0077 0.01142 -0.0050 0.0265 -0.0034 -0.03663 0.0058 0.0713* 0.0133 0.03024 0.0078 0.0452 0.0432 0.01465 -0.0334 0.0142 -0.0424 -0.0345
Notes: The k-order cross-correlation coefficient between two standardised data series x and y is estimated as( ) ∑∑∑ −− = 22
ttkttktt yxyxy,xρ where k represents the number of lags (leads when negatives) of y with
respect x. The standardised residuals in panel (A) are computed as ( )t,t,t,ˆ 111 εσεε = and
( )t,t,t,ˆ 222 εσεε = where σ (.) means the sample standard deviation. The standardised residuals in panel
(B) are computed as t,t,t, hU 1111 ε= and t,t,t, hU 2222 ε= where hii,t represents the conditionalvariance series estimated in Table V. For a sample size of T observations, the asymptotic distribution of the
T times the cross-correlation coefficient is a zero-one normal distribution, that is( ) ( )10,ANy,xT ktt →−ρ (see Cheung and Ng (1996) for more details). Significant coefficients are highlighted
with one (*) and two (**) asterisks are significant at 90% and 95% of confidence level, respectively.
26
Table IX trading rules according to the ‘feedback’ hypothesis on volatility
Panel (A): Trading after bad newsPanel (A.1): ‘Direct’ strategies on the IBEX-35
Strategy Signal in ‘t−1’ Position to take in ‘t−1’A1 [ ] 0011111 ><∆ −− t,tt hE ε Short IBEX-35
B1 [ ] 0011111 <<∆ −− t,tt hE ε Long IBEX-35
Panel (A.2): ‘Crossed’ Strategies on the IBEX-Compl. taking signals from the IBEX-35A2 [ ] 0011111 ><∆ −− t,tt hE ε Short IBEX-Compl.B2 [ ] 0011111 <<∆ −− t,tt hE ε Long IBEX-Compl.
Panel (A.3): ‘Direct’ strategies on the IBEX-Compl.A3 [ ] 0012221 ><∆ −− t,tt hE ε Short IBEX-Compl.B3 [ ] 0012221 <<∆ −− t,tt hE ε Long IBEX-Compl.
Panel (A.4): ‘Crossed’ Strategies on the IBEX-35 taking signals from the IBEX-Compl.A4 [ ] 0012221 ><∆ −− t,tt hE ε Short IBEX-35
B4 [ ] 0012221 <<∆ −− t,tt hE ε Long IBEX-35
Panel (B): Trading after good news
Panel (B.1): ‘Direct’ strategies on the IBEX-35A5 [ ] 0011111 >>∆ −− t,tt hE ε Short IBEX-35B5 [ ] 0011111 <>∆ −− t,tt hE ε Long IBEX-35
Panel (B.2): ‘Crossed’ Strategies on the IBEX-Compl. taking signals from the IBEX-35A6 [ ] 0011111 >>∆ −− t,tt hE ε Short IBEX-Compl.B6 [ ] 0011111 <>∆ −− t,tt hE ε Long IBEX-Compl.
Panel (B.3): ‘Direct’ strategies on the IBEX-Compl.A7 [ ] 0012221 >>∆ −− t,tt hE ε Short IBEX-Compl.B7 [ ] 0012221 <>∆ −− t,tt hE ε Long IBEX-Compl.
Panel (B.4): ‘Crossed’ Strategies on the IBEX-35 taking signals from the IBEX-Compl.A8 [ ] 0012221 >>∆ −− t,tt hE ε Short IBEX-35B8 [ ] 0012221 <>∆ −− t,tt hE ε Long IBEX-35
Where h11t and h22t are the conditional estimates of IBEX-35 variance and IBEX-Complementario variance,respectively, obtained from the asymmetric GARCH model. ε1t-1 is the return shock to the IBEX-35 and ε2t-1
is the return shock to the IBEX-Complementario. And [ ]·Et 1− is the expectation operator conditioned toinformation available in ‘t−1’.
27
Table XEx-post profitability of trading rules according to the ‘feed-back hypothesis’ on
volatilityPeriod: January 2nd, 2001 to June 30th, 2002
Panel (A): Trading after bad newsPanel (A.1): ‘Direct’ strategies on the IBEX-35
Strategy Weeks (+) Weeks (-) Transactions Return (%) TC: 0.1% TC: 0.5% TC: 1% TC: 2%A1 9 8 12 15.06 13.86* 9.06 3.06 -8.94B1 12 10 14 3.99 2.59* -3.01 -10.01 -24.01
Panel (A.2): ‘Crossed’ Strategies on the IBEX-Compl. taking signals from the IBEX-35A2 8 9 12 23.48 22.28 17.48 11.48** -0.52B2 13 9 14 4.49 3.09 -2.51 -9.51 -23.51
Panel (A.3): ‘Direct’ strategies on the IBEX-Compl.A3 7 5 9 26.46 25.56 21.96 17.46** 8.46B3 9 8 10 6.14 5.14 1.14 -3.86 -13.86
Panel (A.4): ‘Crossed’ Strategies on the IBEX-35 taking signals from the IBEX-Compl.A4 9 3 9 33.45 32.55* 28.95 24.45 15.45B4 8 9 10 -2.68 -3.68 -7.68 -12.68 -22.68
Panel (B): Trading after good newsPanel (B.1): ‘Direct’ strategies on the IBEX-35
A5 5 3 7 2.81 2.11* -0.69 -4.19 -11.19B5 12 19 19 -16.75 -18.65 -26.25 -35.75 -54.75
Panel (B.2): ‘Crossed’ Strategies on the IBEX-Compl. taking signals from the IBEX-35A6 3 5 7 -4.60 -5.30 -8.10 -11.60 -18.60B6 20 11 19 14.93 -13.03 -5.43 -4.07 -23.07
Panel (B.3): ‘Direct’ strategies on the IBEX-Compl.A7 6 6 10 -0.11 -1.11 -5.11 -10.11 -20.11B7 27 10 20 20.75 18.75 10.75 0.75** -19.25
Panel (B.4): ‘Crossed’ Strategies on the IBEX-35 taking signals from the IBEX-Compl.A8 8 4 10 12.52 11.52* 7.52 2.52 -7.48B8 20 17 20 18.03 16.03* 8.03 -1.97 -21.97
Panel (C): Buy and hold strategies
IBEX-35 35 43 0 -30.63IBEX-Complementario 47 31 0 0.54
Risk-free 78 0 0 5.95
Notes: The model for means and volatility is estimated for the whole sample and strategies results computedin the period January 2nd, 2001 to June 30th, 2002. In this period, the estimated conditional volatility is takenas ‘forecasted’ values and strategies results computed with the observed stock index values. The remainingcomments can be seen in Table X notes.
28
Table XIEx-ante profitability of trading rules according to the ‘feed-back hypothesis’ on volatility
Period: January 2nd, 2001 to June 30th, 2002
Panel (A): Trading after bad newsPanel (A.1): ‘Direct’ strategies on the IBEX-35
Strategy Weeks (+) Weeks (-) Transactions Return (%) TC: 0.1% TC: 0.5% TC: 1% TC: 2%A1 8 7 11 15.10 14.00* 9.60 4.10 -6.90B1 13 11 16 4.02 2.42* -3.98 -11.98 -27.98
Panel (A.2): ‘Crossed’ Strategies on the IBEX-Compl. taking signals from the IBEX-35A2 7 8 11 24.80 23.70 19.30 13.80** 2.80B2 14 10 16 5.81 4.21 -2.19 -10.19 -26.19
Panel (A.3): ‘Direct’ strategies on the IBEX-Compl.A3 7 7 9 26.05 25.15 21.55 17.05** 8.05B3 6 8 11 0.08 -1.02 -5.42 -10.92 -21.92
Panel (A.4): ‘Crossed’ Strategies on the IBEX-35 taking signals from the IBEX-Compl.A4 9 5 9 24.75 23.85* 20.25 15.75 6.75B4 6 8 11 -10.69 -11.79 -16.19 -21.69 -32.69
Panel (B): Trading after good newsPanel (B.1): ‘Direct’ strategies on the IBEX-35
A5 6 3 8 3.53 2.73* -0.47 -4.47 -12.47B5 12 18 19 -16.02 -17.92 -25.52 -35.02 -54.02
Panel (B.2): ‘Crossed’ Strategies on the IBEX-Compl. taking signals from the IBEX-35A6 3 6 8 -6.05 -6.85 -10.05 -14.05 -22.05B6 19 11 19 13.03 11.13 3.53 -5.97 -24.97
Panel (B.3): ‘Direct’ strategies on the IBEX-Compl.A7 5 9 12 -0.53 -1.73 -6.53 -12.53 -24.53B7 25 11 19 25.98 24.08 15.48 6.98** -12.02
Panel (B.4): ‘Crossed’ Strategies on the IBEX-35 taking signals from the IBEX-Compl.A8 7 7 12 0.64 -0.46 -5.46 -11.46 -23.46B8 17 19 19 5.46 3.56* -4.04 -13.54 -42.54
Panel (C): Buy and hold strategies
IBEX-35 35 43 0 -30.63IBEX-Complementario 47 31 0 0.54
Risk-free 78 0 0 5.95
Notes: The model for means and volatility is estimated each week in the period January 2nd, 2001 to June 30th, 2002.In this period, the conditional covariance matrix is forecasted and trading strategies designed following Table IX, theneach strategy result is computed with the observed stock index value. The ‘Weeks(+)’ (‘Weeks(-)’) column displays thenumber of weeks that each strategy has a positive (negative) return. The ‘Transactions’ column displays the number oftrades by each strategy, taking into account only the weeks the portfolio position changes. Positive returns after takingrealistic transactions costs away are highlighted with one asterisk (*) in the case of strategies on the IBEX-35 and withtwo asterisks (**) on the IBEX-Complementario strategies. Realistic transaction costs are about 0.1% per transaction onthe IBEX-35 but trading with its futures contract and about 1% per transaction in the case of the IBEX-Complementariofor institutional investors.
29
7. Figures
0
2000
4000
6000
8000
10000
12000
14000
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002
Inde
x po
ints
IBEX-35 IBEX-Complementario
Figure 1. Evolution of the stock indices IBEX-35 and IBEX-Complementario
10
15
20
25
30
35
40
45
50
1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002
Ann
ualis
ed c
ondi
tiona
l vol
atili
ty
Ibex-35 Ibex-Complementario
Figure 2. Annualised conditional volatility of the stock indices IBEX-35 and IBEX-
Complementario
30
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002
Cor
rela
tion
Conditional Correlation Unconditional Correlation
Figure 3-a. Unconditional and conditional correlation between the stock indices IBEX-35 and
IBEX-Complementario
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002
Bet
a
Conditional Beta Unconditional Beta
Figure 3-b. Unconditional and conditional Beta coefficients of IBEX-Complementario stock
index with respect the stock index IBEX-35
32
Figure 5-A. A positive shock in the IBEX-Complementario
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F IB
EX-3
5
Figure 5-D. A positive shock in the IBEX-35
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F IB
EX-3
5
Figure 5-B. A positive shock in the IBEX-Complementario
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F C
ovar
ianc
e
Figure 5-E. A positive shock in the IBEX-35
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F C
ovar
ianc
e
Figure 5-C. A positive shock in the IBEX-Complementario
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F IB
EX-C
ompl
emen
tario
Figure 5-F. A positive shock in the IBEX-35
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F IB
EX-C
ompl
emen
tario
Figure 5. Asymmetric volatility impulse-response function to positive unexpected shocks from
the VECM - Asymmetric BEKK
(Dashed lines displays the 90% confidence interval)
33
Figure 6-A. A negative shock in the IBEX-Complementario
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F IB
EX-3
5
Figure 6-D. A negative shock in the IBEX-35
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F IB
EX-3
5
Figure 6-B. A negative shock in the IBEX-Complementario
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F C
ovar
ianc
e
Figure 6-E. A negative shock in the IBEX-35
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F C
ovar
ianc
e
Figure 6-C. A negative shock in the IBEX-Complementario
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F IB
EX-C
ompl
emen
tario
Figure 6-F. A negative shock in the IBEX-35
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
Weeks
VIR
F IB
EX-C
ompl
emen
tario
Figure 6. Asymmetric volatility impulse-response function to negative unexpected shocks
from the VECM - Asymmetric BEKK
(Dashed lines displays the 90% confidence interval)
34
8. References
Badrinath, S. G., Kale, J. Y., and Noe, T. H. (1995). Of Shepherds, Sheep, and the Cross-Autocorrelations in Equity Returns. The Review of Financial Studies, 8, 401-430.
Bekaert, G., and Wu, G. (2000). Asymmetric Volatility and Risk in Equity Markets. The Review ofFinancial Studies, 13, 1-42.
Black, F. (1976). Studies of Stock Price Volatility Changes. Proceedings of the 1976 Meetings ofthe American Statistical Association, Business and Economical Statistics Section, 177-181.
Bollerslev, T., Engle R. F., and Wooldridge, J. M. (1988). A Capital Asset Pricing Model withTime Varying Covariances. Journal of Political Economy, 96, 116-31.
Bollerslev, T. (1990). Modelling the Coherence in Short-Run Nominal Rates: A MultivariateGeneralized ARCH Approach. Review of Economics and Statistics, 72, 498-505.
Bollerslev, T., and Wooldridge, J. M. (1992). Quasi-maximum Likelihood Estimation and Inferencein Models with Time Varying Covariances. Econometric Review, 11, 143-172.
Boudoukh, J., Richardson, M. P., and Whitelaw, R. F. (1994). A Tale of Three Schools: Insights onAutocorrelations of Short-Horizon Stock Returns. The Review of Financial Studies, 7, 539-573.
Braun, P. A., Nelson, D. B., and Sunier, A. M. (1995). Good News, Bad News, Volatility andBetas. The Journal of Finance, 5, 1575-603.
Brenner, R. J., Harjes, R. H. & Kroner, K. F. (1996). Another Look at Models of the Short-TermInterest. Journal of Financial and Quantitative Analysis, 31, 85-107.
Campbell, J. Y., and Henstschel, L. (1992). No News is Good News: An Asymmetric Model ofChanging Volatility in Stock Returns, Journal of Financial Economics, 31, 281-318.
Chelley-Steeley, P. L., and Steeley, J. M. (1995). Conditional Volatility and Firm Size: AnEmpirical Analysis of UK Equity Porfolios. Applied Financial Economics, 5, 433-440.
Cheung, Y.W., and Ng, L. K. (1996). A Causality-in-Variance Test and its Application to FinancialMarket Prices. Journal of Econometrics, 72, 33-48.
Chopra, N., Lakonishok, J., and Ritter, J. R. (1992). Measuring Abnormal Performance. Do StocksOverreact?. Journal of Financial Economics, 31, 235-268.
Christie, A. A. (1982). The Stochastic Behavior of Common Stock Variances: Value, Leverage andInterest Rate Effects. Journal of Financial Economics, 10, 407- 432.
Conrad, J., Gultekin, M. N., and Kaul, G. (1991). Asymmetric Predictability of ConditionalVariances. The Review of Financial Studies, 4, 597-622.
Conrad, J., Kaul, G., and Nimalendran, M. (1991). Components of Short-Horizon IndividualSecurity Returns. Journal Financial Economics, 29, 365-384.
Dickey D. A., and Fuller, W.A. (1981). Likelihood Ratio Statistics for Autoregressive Time Serieswith a Unit Root. Econometrica, 49, 1057-72.
Enders, W. (1995). Applied Econometric Time Series. New York: John Wiley and Sons.Engle, R. F., and Granger, C. W. (1987). Cointegration and Error Correction: Representation,
Estimation and Testing. Econometrica, 55, 251-276.Engle, R. F., and Kroner, K. F. (1995). Multivariate Simultaneous Generalized ARCH. Econometric
Theory, 11, 122-50.Engle, R. F., and Ng, V. K. (1993). Measuring and Testing the Impact of News on Volatilily. The
Journal of Finance, 48, 1749 – 78.Engle, R. F., Ng, V. K., and Rothschild, M. (1990). Asset Pricing with a Factor-ARCH Covariance
Structure: Empirical Estimates for Treasury Bills. Journal of Econometrics, 45, 213-38.Glosten, L. R., Jagannathan, R., and Runkel, D. E. (1993). On the Relation between the Expected
Value and Volatility of Nominal Excess Return on Stocks. The Journal of Finance, 48,1779-801.
Hamilton, J. D. (1994). Time Series Analysis. Princeton, New Jersey: Princeton University Press.Hentschel, L. (1995). All in the Family: Nesting Symmetric and Asymmetric GARCH Models.
Journal of Financial Economics, 39, 71-104.
35
Johansen, S. (1995). Likelihood-Based Inference in Cointegrated Vector Auto-Regressive Models.Oxford University Press.
Johansen, S. (1988). Statistical Analysis of Cointegration Vectors. Journal of Economic Dynamicsand Control, 12, 231-254.
Kroner, K. F., and Ng, V. K. (1998). Modeling Asymmetric Comovements of Asset Returns. TheReview of Financial Studies, 11, 817-844.
Lin, W.L. (1997). Impulse Response Function for Conditional Volatility in GARCH Models.Journal of Business and Economic Statistics, 15, 15-25.
Lo, A. W., and MacKinlay, C. (1990). When are Contrarian Profits Due to Stock MarketOverreaction? Review of Financial Studies, 3, 175-205.
MacKinnon, J. (1991). Critical Values for Co-integration Tests. In Engle, R. F., and Granger, C. W.L. (eds). Long-Run Economic Relationships. Oxford University Press, 767-774.
McQueen, G., Pinegar, M., and Thorley, S. (1996). Delayed Reaction to Good News and the Cross-Autocorrelation of Porfolio Returns. Journal of Finance, 51, 889-919.
Meneu, V., and Pardo, A. (2003). Pre-holiday Effect, Large Trades and Small Investor Behaviour.Forthcoming in the Journal of Empirical Finance.
Meneu, V., and Torro, H. (2003). Asymmetric Covariance in Spot-Futures Markets. The Journal ofFutures Markets, Forthcoming November 2003.
Nam, K., Pyun, C. S., and Avard, S. L. (2001). Asymmetric Reverting Behavior of Short-HorizonStock Returns: An Evidence of Stock Market Overreaction. Journal of Banking and Finance,25, 807-824.
Olson, D., and Mossman, K. (2001). Cross-correlations and Predictibility of Stock Returns. Journalof Forecasting, 20, 145-160.
Osterwald-Lenum, M. (1992). A Note with Quantiles of the Asymptotic Distribution of the MLCointegration Rank Test Statistics. Oxford Bulletin of Economics and Statistics, 50, 361-377.
Phillips, P. C. B., and Perron, P. (1988). Testing for a Unit Root in Time Series Regression.Biometrica, 75, 335-346.
Ross, S. A. (1989). Information and Volatility: The No-Arbitrage Martingale Approach to Timingand Resolution Irrelevancy. The Journal of Finance, 44, 1-17.
Schwert, G. W. (1989). Why Does Stock Market Volatility Change Over Time? The Journal ofFinance, 44, 1115-1153.
Veronesi, P. (1999). Stock Market Overreaction to Bad News in Good Times: A RationalExpectetions Equilibrium Model. The Review of Financial Studies, 12, 975-1007.
Wooldridge, J. M. (1990). A Unified Approach to Robust, Regression-based Specification Tests.Econometric Theory, 6, 17-43.
Wu, G. (2001). The Determinants of Asymmetric Volatility. The Review of Financial Studies, 14,837-859.
Yu, C.-H., and Wu, C. (2001). Economic Sources of Asymmetric Cross-Correlation among StockReturns. International Review of Economics and Finance, 10, 19-40.